Mixed emotions about linear mixed models

[Erik Langer Madsen]

I am currently investigating a dataset with 68 subjects defined by the
fixed factors gender:m/f,  diabetes: yes/no treatment: active/placebo and
the covariate age and possibly the random factor clinic-no 1, 2 or 3.
The subjects attended a weight loss trial at three different clinics and
had various parameters including their weight measured at 5 timepoints
over three years.

Due to the repeated data being correlated and due to some missing values
at some timepoints for some subjects linear mixed seemed to be the perfect
choice. However I'm being mentally disoriented (pronunciation: scratching
my head and dont know what to make of it) due to the following problems

1 which covariance structure should I aim for? At present I'm using ar1
but would unstructured or compound symmetry be more appropriate? How to
validate the model components (Wald ?)

2 Adding clinics as a random factor generates a non positive hessian
matrix although convergence criteria are fulfilled. Ignoring this warning
doesn't feel appropriate. Any explanations or suggestions are very welcome.

3 The p-values and estimated marginal means changes a lot when gradually
increasing the number of fixed factors in the model. And the validity of
the p-values is beginning to sound a little hollow when they change from
model to model. Any explanations or suggestions are very welcome.

4. Entering subject into the random factors as a variance component doesnt
seem to change anything if you paste the syntax with or without this
action. Is it irrelevant or ?

Hopefully somebody else has experienced similar problems otherwise I'm
completely ready to admit my mental or experimental shortcomings in return
for some advice

Best regards
Erik
Phd-student MD
Aarhus Denmark

[MaxJasper]

Wald test is not accurate. So you should compare several models based on
their -2RR differences and then select check to see if you like a model
based on whether -2RR (model 1 - model 2) with Chi-square distribution and
df=(parameters in model 1 - parameter in model 2) is significant or not.



|
|1 which covariance structure should I aim for? At present I'm
|using ar1 but would unstructured or compound symmetry be more
|appropriate? How to validate the model components (Wald ?)
|

[Erik Langer Madsen]

On Wed, 11 Jul 2007 10:41:43 -0600, Max Jasper <[hidden email]> wrote:

>Wald test is not accurate. So you should compare several models based on
>their -2RR differences and then select check to see if you like a model
>based on whether -2RR (model 1 - model 2) with Chi-square distribution and
>df=(parameters in model 1 - parameter in model 2) is significant or not.
>
>
>
>|
>|1 which covariance structure should I aim for? At present I'm
>|using ar1 but would unstructured or compound symmetry be more
>|appropriate? How to validate the model components (Wald ?)
>|

Dear maxjasper

Thank you for using your time to evaluate my problem. I'll look into your
suggestions and see if anything works out

best regards
Erik